A smart, practical, deep learning-based clinical decision support tool for patients in the prostate-specific antigen...

Area of Science:

• Urology
• Oncology
• Artificial Intelligence in Medicine

Background:

• Prostate cancer (PC) screening and early detection face challenges, particularly in patients with intermediate prostate-specific antigen (PSA) levels (PSA gray zones).
• Existing biomarkers show limited accuracy in diagnosing PC and clinically significant PC (CSPC) within these PSA gray zones.

Purpose of the Study:

• To develop and validate a deep learning-based prediction model for PC and CSPC.
• The model aims to utilize minimized parameters and incorporate missing value handling algorithms for improved diagnostic performance.

Main Methods:

• Retrospective analysis of 12,739 prostate biopsy cases within the PSA gray zone (2.0-10.0 ng/mL).
• Development of Dense Neural Network (DNN) and Extreme Gradient Boosting (XGBoost) models using 5-fold cross-validation.
• Comparison of model performance (AUROC) against serum PSA, PSA density, free PSA (fPSA), and Prostate Health Index (PHI).

Main Results:

• The DNN model demonstrated strong predictive performance for both PC and CSPC, outperforming traditional clinical biomarkers.
• Internal and external validation showed competitive AUROC values for the DNN model, even with missing value imputation.
• The DNN model's performance was comparable to XGBoost and superior to PHI, serum PSA, and percent-fPSA.

Conclusions:

• Deep learning models, particularly DNNs with missing value imputation, show significant potential for predicting PC and CSPC in PSA gray zones.
• While further real-world validation is necessary, these findings support the growing utility of deep learning in clinical diagnostics.
• The developed model offers a promising tool for enhancing prostate cancer diagnosis using routinely available clinical parameters.